AI/LLM engineer (LatAm)

MincaAI builds AI-powered workflow automation software for the insurance industry. We design intelligent agents that automate complex, repetitive operational processes — from document intake and data extraction to smart matching and decision support — so that teams can focus on high-value work instead of manual tasks.

We are looking for an AI/LLM Engineer to own the LLM layer across our platform: what it extracts, what it classifies, how we know it works, what it costs, and how it improves every week without a retrain.

🎯 The problem

Insurance runs on documents that were never designed to be read by a machine. A broker sends a fleet spreadsheet with headers in three conventions. A carrier returns a quotation as a PDF. An endorsement arrives as an email where the number that matters sits inside a sentence.

All of it has to become structured, typed, validated data — with a confidence score on each field, because a field that is silently wrong costs more than a field that is missing.

We already run this in production: our vehicle codifier resolves free-text descriptions against the Mexican AMIS catalog at ~84% Top-1, with a three-way decision and a review queue that feeds corrections back in. Extraction and classification sit around it, and that is where most of the remaining error budget lives.

🔧 What you'll do

• Own extraction end-to-end — spreadsheets, PDFs, emails, OCR output into typed records with per-field confidence. Schema, prompts, validation, repair.
• Build classification components — document type, vehicle type, coverage, use type. Measured models with owned metrics, not prompt strings someone tweaks.
• Own the evaluation harness — golden sets, regression suites, CI gates, drift tooling between prompt and model versions.
• Instrument and optimise the AI layer — quality, latency, degradation, cost per call. Model routing, caching, structured outputs, provider fallback.
• Close the feedback loop — reviewer corrections become few-shot banks, alias dictionaries, routing rules, eval cases.
• Work with the client's underwriters — build ground truth with Spanish-speaking domain experts and defend thresholds to the people who act on them.

🏗️ Our stack

• Backend — Python 3.14, FastAPI, gRPC, Pydantic, SQLAlchemy 2, Celery, Redis
• Data — PostgreSQL + pgvector (HNSW, GIN, trigram), S3
• AI/ML — OpenAI and Gemini with provider fallback, embeddings, hybrid retrieval, scikit-learn
• Infra — Docker, Kubernetes, AWS, GitHub Actions, structlog + OpenTelemetry
• Quality — strict mypy, ruff, black, unit / integration / live-LLM test tiers

✅ What we're looking for

• 5+ years of professional software engineering, including 2+ shipping LLM systems that others depend on in production
• Expert-level Python: typed, tested, maintainable. FastAPI or equivalent
• Production LLM engineering: structured outputs, tool calling, schema validation and repair, provider fallback, caching, latency and cost budgets
• Prompts treated as versioned, tested artifacts — a diff, a test and a number for every change you shipped
• Information extraction and classification at production quality, out of noisy and semi-structured sources
• Confidence estimated per field, not per document
• Evaluation you built yourself, not just consumed: golden sets, labelling protocols, held-out discipline, regression detection in CI
• Error analysis that attributes a two-point drop to a stage and a cause
• Comfort with the whole toolbox — LLM, fine-tuned small model, classical ML, deterministic rule — chosen on evidence, not preference
• PostgreSQL you can reason about: query plans, indexes, JSONB
• Strong ownership mentality; you don't wait for instructions, you drive solutions
• Comfortable making architectural decisions and defending technical trade-offs
• Experience in startups or small autonomous teams
• Spanish and English required. The data is in Spanish and you cannot debug what you cannot read

🎚️ The bar

We are hiring one person for this and we would rather leave the seat empty than lower it.

• You have shipped an LLM system you were measured on. Not a prototype — a system in production whose quality number was yours.
• Evaluation is a reflex. You have built a golden set, argued about what belongs in it, caught a regression with it, and abandoned one of your own ideas because the numbers said so. If your instinct on hearing "accuracy dropped two points" is to ask which stage lost it — we want to talk. This is the single most important line on this page. Everything else is negotiable. This is not.
• You know what an LLM is bad at, and you say so. The model is one component among several. If your answer to every extraction problem is a better prompt, this will frustrate you and you will frustrate us.
• Comfort with ambiguity. Nobody will hand you a spec for what the system should output when the input itself is ambiguous. You will write that spec.

If the evaluation line describes you exactly, apply even if something else on this page does not.

📋 Details

• Contract: Freelance
• Duration: 4 months (extendable)
• Location: 100% Remote (LATAM)

📩 Send your CV to [email protected]